[Martin Taylor 970305 13:40]
Rick Marken (970304.0840 PST)]
Martin Taylor (970303 12:20) --
> Bruce's question [Control is the _only_ problem] is much better
> than "Determining the controlled variable is the only problem."
I agree that Bruce Gregory's statement of the problem is excellent. But
I don't think it's fair to characterize my position as simply:
"Determining the controlled variable is the only problem."
All right. I apologize. But in my defence, I think it only fair to say
that the words you use often sound an awful lot like "Determining the
controlled variable is the only problem."
What you say in the following is very sensible. As is a lot of what you
write, when not ranting.
I think it
would be better to say that my position has been that "Determining
the controlled variable must be part of any research aimed at
studying control". If you look at your excellent list of research
problems, for example, I think you will see that every one of them
involves or presupposes the determination of controlled variables.
Perhaps my position could be made clearer by casting it in terms of a
concrete example: the study of the behavior of a thermostat. There
are many things that you might want to find out about the behavior of a
thermostat (or of interacting thermostats). You might want to know if a
thermostat shifts control from one variable to another, whether its
ability to control affected by experience (learning), under what
conditions it shifts to control of an imagined perception, what kind of
conditions lead to collaboration and what kind lead to conflict between
thermostats, etc. But, as you can see, all of these problems can be
addressed productively _only_ if you know what variable(s) the
thermostat is controlling. If you don't know that the thermosat is
controlling sensed temperature (rather than humidity, pressure,
acoustical vibrations, light level, or nothing at all) then you can't
answer any questions about the behavior of the thermostat sensibly.
I think the problem arises when you see research being done without
regard to the fact that the behaviour an experimenter sees must be in some
way aimed at controlling a perceptual variable. If an experimenter
doesn't realize that it is, he or she is likely to make all sorts
of invalid inferences. The research may be fine in one way--that if
it is done again under similar circumstances, similar results will
probably be obtained--but absolutely lousy in the important way--that
it leads to inappropriate interpretations of how people work. And to
wrong applications in the real world for which the Laboratory is supposed
to provide a simplified model.
When you see words that you can stretch so that they might conceivable seem
to support such research, whether or not they were so intended, it
seems to lead to what often looks like an S-R "behaviour." On go
the blinkers, and out come extreme and unsupportable statements. And I
mean unsupportable within the context of PCT. But not today...
I think conventional psychology is in the position of the person who
wants to study all kinds of interesting things about the behavior of
a thermostat but has no idea 1) that the thermostat is a control system
and 2) that it controls a perceptual reprentation of some environmental
state of affairs (temperature). This is why I think
it's important for conventional psychologists to understand the
_possibility_ that the systems they study are control systems and
that the _only_ way to find out whether or not these systems
actually _are_ control systems (and, if so, what they are
controlling) is by Testing for Controlled Variables.
Fair enough, like the rest of this particular posting. But having got
that far, there is a lot further to go on the way to understanding
human behaviour.
···
----------------------
Now I'd like to consider a previous posting I'd class as "sensible but
overlooks a point."
Rick Marken (970302.1700 PST) to Bruce Abbott:
The problem is that you can't determine the open-loop characteristics of
these "open-loop" systems while they are part of a closed loop.
You can, and the reason your argument fails is that you forget that the
single scalar variables in the formulae that describe the loop are in many
cases functions of several variables. The output affects the reference
signals of several lower-level feedback loops, each muscular influence
on the environment affects many physical factors, each low-level perception
is a function of many sensory variables, and the single controlled
perception that defines the loop is a function of many lower-level
perceptions.
One can break any of the lower level loops without necessarily affecting
the ability to control the one that defines the loop you are talking
about. For example, in normal life, one may choose to raise or lower
the thermostat setting if one is too warm or too cool. In an experiment
on judging temperature, the experimenter prevents the subject from doing
that. But the subject can still control for seeing the experimenter to
be satisfied with the subject's performance. To control for that, the
subject is permitted the output of indicating (with a word or a button
push) which way she would have changed the thermostat had the perception
of temperature been controllable.
To control temperature, differences of temperature must be perceptible.
The experimenter ensures that no action of the subject
can affect the perceived temperature, but that doesn't stop the
subject from perceiving it and on using that perception in some way
to affect the reference values for other perceptions (with control
outputs that lead to the experimenter seeing button pushes or hearing
words).
You can clamp of break some of the strands in a feedback loop quite readily,
without affecting the ability to control those things the feedback loop
might have been supporting. It happens all the time in real life that
sometimes you can control a particular perception and at other times
you can't, and have to find another way to control the higher-level
perceptions. Remember the aphorism "Many means to the same end."
So although you may be right in your paragraph:
The relationship p = f(q.i), for example, exists in a closed loop system
where it is _simultaneously_ true that p = f(q.i) and
q.i = g(r-p)+d. In this situation it is impossible to _manipulate_
q.i as an _independent variable_ in order to determine the nature
of the relationship, f(), between q.i and p. This is because q.i
is also a dependent variable -- indeed, it is ultimately dependent
on itself via the closed loop. If the feedback in the closed loop is
negative then q.i is a _controlled variable_ making it mathematically
and _physically_ impossible to treat q.i as an indenpedent variable.
Nevertheless you are wrong in saying that you cannot make "g" equal to
zero in q.i = g(r-p)+d.
When you make g = 0, you clearly cannot determine p = f(q.i) from its
effect on q.o (= g(r-p)). So you have to work on the fact that p enters
into some other control loop. And to discover/produce/guess-at/force
a control loop that has as a component an isolatable relationship between
p and some component of its output is the task of a good experimenter
looking for the open-loop characteristics of p=f(q.i).
Of cousre,
it is necessary to know what variable the system is controlling
(that is, it is necessary to know what aspect of the environment
corresponds to q.i) before one can start modeling the open loop
characteristics of the functional relationships in the control loop.
I'm not at all sure about this. I'll go along with it being much easier,
and perhaps much more reliable to determine what is perceived if one
does the Test properly. But "necessary" is what I'm not sure about.
Here's an actual scenario, not a hypothetical one, though the detail may
have got changed in my memory. Around 1960, a founder of signal detection
theory in psychology (W.P.Tanner) told me about a study he had done.
I don't know if it was published.
Tanner's idea was that people learned to perceive particular things
only because the environment gave them reliable feedback for
differentially perceiving changes in the sensory patterns. So he
looked for a sensory pattern to which nobody had ever been exposed, and
tried to see whether he could get people to perceive it reliably. He
didn't use "the Test", but used ordinary psychophysical methods, telling
the subjects what had been the correct answer after each discrete trial.
The sensory pattern was like this. In one ear he played a waveform (sine
or noise--I don't remember) for which the amplitude started at a maximum
and declined linearly over a duration of 100 msec (or so). In the other
ear he played the same waveform, but with an amplitude that ramped up
instead of down, so that the total power summed across the earphones
stayed constant over the 100 msec double-ramp. At some moment in the
ramp he switched all the power into one ear, cutting off the input to
the other ear for a short period (20 msec?). This moment might come early
or late in the ramp (at four times, if I remember correctly). All the
way through, the total power delivered to the two ears stayed constant.
The subject was supposed to report at which of the four times the short
monaural pulse happened. It's an almost impossible task for someone who
hasn't learned how to hear this signal--a signal that cannot occur in the
natural world. And at first, nobody got more than chance performance (d'=0
in detection-theory terms). But Tanner got his subjects to keep at it,
and every one of them learned to get a perfect score eventually. In all
cases, the shift from being quite unable to hear the differences to being
essentially perfect happened quite quickly--within one or two experimental
sessions, I think--even though for at least one subject it took 45 days
of being unable to hear anything before this "Aha" perception occurred.
Tanner's question was "can people learn to hear fine discriminations in
something they simply could never have heard before and could initially
not discriminate at all, using a signal in which the _physical_ discrimination
is quite large." Had he applied "the Test", his answer would have been
"People don't control that perception." This would have been an answer to
a different question.
Tanner didn't ask "Can people learn to control a perception corresponding
to this physical discrimination," but his results seem to be immediately
applicable to that question. The answer is "Yes." Once they had learned
to make the discrimination in an open-loop study, they would almost certainly
be able to use it if they were given a control knob to shift the timing
of the pulse. Before they had learned it, they probably would not have
been able to do so.
Tanner was interested in the capabilities of the auditory system, rather
in how it is ordinarily used. The Test can be used for the same purpose,
but as ordinarily construed, it asks the question "What variable is the
subject controlling" rather than "Can the subject learn to control function
X." It's quite possible that Tanner's subjects would have taken a lot
less time to learn the discrimination if they had been given control of
the timing, and been provided with, say, a cursor tied to the actual
timing, which they had to use to track a target on a screen. But he did
use feedback in a loop of which the "interesting" perception was a
component, and he used it because of a firm belief that the feedback
was an essential aspect of learning perceptions in everyday life.
Open loop studies _can_ be performed on people, who are nevertheless
acting as control systems at all times. And they can provide useful
results.
That's the bottom line. (Of this posting).
(No, This is the bottom line, sorry).
Martin
